A brain-inspired-AI researcher's other job: co-founding China's first cross-institutional AI safety body, which a 2026 index still grades an F

Core claims

1. CnAISDA (Beijing-AISI's parent body), launched Feb 2025, encodes its safety priority differently for each audience in its own bilingual name. "In the English name, 'safety' comes before 'development,' while in Mandarin the order is reversed. This subtle difference may reflect an effort to emphasize safety concerns when engaging with international audiences while maintaining development as the primary focus domestically." — Carnegie Endowment (Singer, Elmgren, Guest), Tier 2.

2. Turing Award laureate Andrew Yao, one of CnAISDA's founders, frames the stakes in extinction terms: "We have suddenly found a way to create a new species that is many, many times more powerful than we are... if we don't do anything, we are going to be eliminated." — same source, Tier 2.

3. A 2026 independent safety index still grades Chinese frontier labs far below Western peers despite this institutional effort: DeepSeek scored 0.47 (F), Alibaba Cloud 0.87 (D-), Z.ai 0.88 (D-), versus Anthropic 2.66 (C+) and OpenAI 2.28 (C). Panelists said deferring wholly to government regulation "amounts to 'complete passivity' as an existential-safety strategy." — Future of Life Institute AI Safety Index, Summer 2026, Tier 2.

Why this was hop-worthy

The entry point (NeuroCogMap, a neuroscience-parcellation framework for reading LLM internals) led to a co-author whose day job — brain-inspired spiking-neural-network research — sits right next to a second, unrelated career as a national AI-safety-governance architect, bridging the vault's technical-AI thread into international policy for the first time.

Further leads

• Andrew Yao's own pivot from 1980s computational-complexity theory to AI-extinction warnings — a possible cross-time-period bridge, not pursued this chain.
• DeepSeek-R1's Jan 2025 release as the stated catalyst tightening China's development-vs-safety tension — not pursued.

Hop chain

Hop 1 — arXiv q-bio.NC/recent listing → "NeuroCogMap Reveals Cognitive Organization of Large Language Models," https://arxiv.org/abs/2607.00397

• Hook type: cross-domain bridge
• Hook: a cognitive-neuroscience method (functional brain parcellation) applied to LLM internals, and shown to improve prediction of real human cortical fMRI responses ("strongest correspondence in higher-order association cortex")
• Why followed: matches the vault's own recurring brain/AI-plausibility thread (Hinton/NGRAD, feedback alignment) but from the opposite direction — neuroscience methods reading AI, not AI modeling the brain
• Key findings: the paper's internal "parcels" are stable across models, link to specific LLM failure modes (hallucination, sycophancy, refusal failure), and predict real cortical activity during language comprehension

Hop 2 — NeuroCogMap co-author search → Yi Zeng institutional bio (LCFI, AI for Good, Turing Institute profile pages)

• Hook type: the person behind the thing
• Hook: a brain-inspired-AI lab director who is also founding dean of Beijing-AISI and a member of the UN High-Level Advisory Body on AI
• Why followed: unfamiliar-name/person hook with no existing claim-note; the dual career (technical neuro-AI + global governance) looked like it might bridge two very different domains
• Key findings: Zeng directs both the Brain-inspired Cognitive AI Lab and the International Research Center for AI Ethics and Governance at CAS, and has briefed international bodies on AI risk

Hop 3 — Yi Zeng → Carnegie Endowment, "How Some of China's Top AI Thinkers Built Their Own AI Safety Institute" (Singer, Elmgren, Guest, June 16 2025), https://carnegieendowment.org/research/2025/06/how-some-of-chinas-top-ai-thinkers-built-their-own-ai-safety-institute

• Hook type: surprising claim
• Hook: the bilingual name-order reversal, and Andrew Yao's extinction-risk quote from a Turing Award-winning theorist
• Why followed: a genuinely new domain for the vault (AI geopolitics/governance), with a culturally resonant detail (deliberate bilingual framing) and a strong quote
• Key findings: CnAISDA is a networked body (integrating Tsinghua, BAAI, ministry centers) rather than new bureaucracy, launched Feb 2025 in Paris, "established with government support" per co-founder Fu Ying, with deliberately ambiguous state ties

Hop 4 — CnAISDA/Beijing-AISI → Future of Life Institute, AI Safety Index Summer 2026, https://futureoflife.org/ai-safety-index-summer-2026/

• Hook type: surprising claim (quantitative)
• Hook: despite the institutional safety apparatus, China's frontier labs score dramatically worse than Western labs on independent safety grading
• Why followed: closes the loop with a quantitative, falsifiable check on whether the institution-building translated into measured outcomes
• Key findings: DeepSeek F (0.47), Alibaba Cloud and Z.ai both D- (0.87/0.88), vs. Anthropic C+ (2.66) and OpenAI C (2.28); report frames passivity toward state regulation as itself a safety failure mode, and notes "inadequate safety is a global problem, not a regional one"

Saved hooks not followed:

• "Shunting Inhibition and Dendritic Branching Shape Local Credit Assignment" (arXiv:2607.03556) — from the same seed listing — reason saved: directly extends the vault's large brain-plausibility-of-backprop cluster (mechanism hook), but the chain went toward the governance bridge instead.
• "Compensation geometry" in diffusion-learning parameter manifolds (arXiv:2607.03671) — from the seed listing — reason saved: unfamiliar-term hook, lower priority than the cross-domain bridge that won.
• Andrew Yao's 1980s complexity-theory background as a cross-time-period bridge to his 2026 extinction warnings — from the Carnegie piece — reason saved: a second person-hook in one chain felt like one too many; good candidate for its own chain.

post-worthy: maybe — well-sourced (Tier 2 throughout, one Tier 1 arXiv anchor) and a genuine new domain (AI geopolitics) bridging into the vault's existing AI-safety-adjacent notes, but it has no MOC home yet; worth a second chain before deciding if it seeds a cluster.